ICML 2018 Papers

Autoregressive sequence models based on deep neural networks, such as RNNs, Wavenet and Transformer are the state-of-the-art on many tasks. However, they lack parallelism and are thus slow for long sequences. RNNs lack parallelism both during training and decoding, while architectures like WaveNet and Transformer are much more parallel during training, but still lack parallelism during decoding. We present a method to extend sequence models using discrete latent variables that makes decoding much more parallel. The main idea behind this approach is to first autoencode the target sequence into a shorter discrete latent sequence, which is generated autoregressively, and finally decode the full sequence from this shorter latent sequence in a parallel manner. To this end, we introduce a new method for constructing discrete latent variables and compare it with previously introduced methods. Finally, we verify that our model works on the task of neural machine translation, where our models are an order of magnitude faster than comparable autoregressive models and, while lower in BLEU than purely autoregressive models, better than previously proposed non-autogregressive translation.

We describe two efficient, and exact, algorithms for computing Bellman updates in robust Markov decision processes (MDPs). The first algorithm uses a homotopy continuation method to compute updates for L1-constrained s,a-rectangular ambiguity sets. It runs in quasi-linear time for plain L1-norms and also generalizes to weighted L1-norms. The second algorithm uses bisection to compute updates for robust MDPs with s-rectangular ambiguity sets. This algorithm, when combined with the homotopy method, also has a quasi-linear runtime. Unlike previous methods, our algorithms compute the primal solution in addition to the optimal objective value, which makes them useful in policy iteration methods. Our experimental results indicate that the proposed methods are over 1,000 times faster than Gurobi, a state-of-the-art commercial optimization package, for small instances, and the performance gap grows considerably with problem size.

Spectral clustering is a widely studied problem, yet its complexity is prohibitive for dynamic graphs of even modest size. We claim that it is possible to reuse information of past cluster assignments to expedite computation. Our approach builds on a recent idea of sidestepping the main bottleneck of spectral clustering, i.e., computing the graph eigenvectors, by a polynomial-based randomized sketching technique. We show that the proposed algorithm achieves clustering assignments with quality approximating that of spectral clustering and that it can yield significant complexity benefits when the graph dynamics are appropriately bounded. In our experiments, our method clusters 30k node graphs 3.9$\times$ faster in average and deviates from the correct assignment by less than 0.1%.

Machine learning models (e.g., speech recognizers) trained on average loss suffer from representation disparity—minority groups (e.g., non-native speakers) carry less weight in the training objective, and thus tend to suffer higher loss. Worse, as model accuracy affects user retention, a minority group can shrink over time. In this paper, we first show that the status quo of empirical risk minimization (ERM) amplifies representation disparity over time, which can even turn initially fair models unfair. To mitigate this, we develop an approach based on distributionally robust optimization (DRO), which minimizes the worst case risk over all distributions close to the empirical distribution. We prove that this approach controls the risk of the minority group at each time step, in the spirit of Rawlsian distributive justice, while remaining oblivious to the identity of the groups. We demonstrate that DRO prevents disparity amplification on examples where ERM fails, and show improvements in minority group user satisfaction in a real-world text autocomplete task.

Sampling methods that choose a subset of the data proportional to its diversity in the feature space are popular for data summarization. However, recent studies have noted the occurrence of bias {–} e.g., under or over representation of a particular gender or ethnicity {–} in such data summarization methods. In this paper we initiate a study of the problem of outputting a diverse and fair summary of a given dataset. We work with a well-studied determinantal measure of diversity and corresponding distributions (DPPs) and present a framework that allows us to incorporate a general class of fairness constraints into such distributions. Designing efficient algorithms to sample from these constrained determinantal distributions, however, suffers from a complexity barrier; we present a fast sampler that is provably good when the input vectors satisfy a natural property. Our empirical results on both real-world and synthetic datasets show that the diversity of the samples produced by adding fairness constraints is not too far from the unconstrained case.

We consider the setting where we wish to perform ranking for hundreds of thousands of users which is common in recommender systems and web search ranking. Learning a single ranking function is unlikely to capture the variability across all users while learning a ranking function for each person is time-consuming and requires large amounts of data from each user. To address this situation, we propose a Factorization RankSVM algorithm which learns a series of k basic ranking functions and then constructs for each user a local ranking function that is a combination of them. We develop a fast algorithm to reduce the time complexity of gradient descent solver by exploiting the low-rank structure, and the resulting algorithm is much faster than existing methods. Furthermore, we prove that the generalization error of the proposed method can be significantly better than training individual RankSVMs. Finally, we present some interesting patterns in the principal ranking functions learned by our algorithms.

We present a novel algorithm that uses exact learning and abstraction to extract a deterministic finite automaton describing the state dynamics of a given trained RNN. We do this using Angluin’s \lstar algorithm as a learner and the trained RNN as an oracle. Our technique efficiently extracts accurate automata from trained RNNs, even when the state vectors are large and require fine differentiation.

The past few years have witnessed growth in the computational requirements for training deep convolutional neural networks. Current approaches parallelize training onto multiple devices by applying a single parallelization strategy (e.g., data or model parallelism) to all layers in a network. Although easy to reason about, these approaches result in suboptimal runtime performance in large-scale distributed training, since different layers in a network may prefer different parallelization strategies. In this paper, we propose layer-wise parallelism that allows each layer in a network to use an individual parallelization strategy. We jointly optimize how each layer is parallelized by solving a graph search problem. Our evaluation shows that layer-wise parallelism outperforms state-of-the-art approaches by increasing training throughput, reducing communication costs, achieving better scalability to multiple GPUs, while maintaining original network accuracy.

Researchers have applied deep neural networks to image restoration tasks, in which they proposed various network architectures, loss functions, and training methods. In particular, adversarial training, which is employed in recent studies, seems to be a key ingredient to success. In this paper, we show that simple convolutional autoencoders (CAEs) built upon only standard network components, i.e., convolutional layers and skip connections, can outperform the state-of-the-art methods which employ adversarial training and sophisticated loss functions. The secret is to search for good architectures using an evolutionary algorithm. All we did was to train the optimized CAEs by minimizing the l2 loss between reconstructed images and their ground truths using the ADAM optimizer. Our experimental results show that this approach achieves 27.8 dB peak signal to noise ratio (PSNR) on the CelebA dataset and 33.3 dB on the SVHN dataset, compared to 22.8 dB and 19.0 dB provided by the former state-of-the-art methods, respectively.

In inductive transfer learning, fine-tuning pre-trained convolutional networks substantially outperforms training from scratch. When using fine-tuning, the underlying assumption is that the pre-trained model extracts generic features, which are at least partially relevant for solving the target task, but would be difficult to extract from the limited amount of data available on the target task. However, besides the initialization with the pre-trained model and the early stopping, there is no mechanism in fine-tuning for retaining the features learned on the source task. In this paper, we investigate several regularization schemes that explicitly promote the similarity of the final solution with the initial model. We show the benefit of having an explicit inductive bias towards the initial model, and we eventually recommend a simple $L^2$ penalty with the pre-trained model being a reference as the baseline of penalty for transfer learning tasks.

This paper studies the estimation of low-rank Markov chains from empirical trajectories. We propose a non-convex estimator based on rank-constrained likelihood maximization. Statistical upper bounds are provided for the Kullback-Leiber divergence and the $\ell_2$ risk between the estimator and the true transition matrix. The estimator reveals a compressed state space of the Markov chain. We also develop a novel DC (difference of convex function) programming algorithm to tackle the rank-constrained non-smooth optimization problem. Convergence results are established. Experiments show that the proposed estimator achieves better empirical performance than other popular approaches.

Training neural networks involves finding minima of a high-dimensional non-convex loss function. Relaxing from linear interpolations, we construct continuous paths between minima of recent neural network architectures on CIFAR10 and CIFAR100. Surprisingly, the paths are essentially flat in both the training and test landscapes. This implies that minima are perhaps best seen as points on a single connected manifold of low loss, rather than as the bottoms of distinct valleys.

We analyze the variance of stochastic gradients along negative curvature directions in certain non-convex machine learning models and show that stochastic gradients indeed exhibit a strong component along these directions. Furthermore, we show that - contrary to the case of isotropic noise - this variance is proportional to the magnitude of the corresponding eigenvalues and not decreasing in the dimensionality. Based upon this bservation we propose a new assumption under which we show that the injection of explicit, isotropic noise usually applied to make gradient descent escape saddle points can successfully be replaced by a simple SGD step. Additionally - and under the same condition - we derive the first convergence rate for plain SGD to a second-order stationary point in a number of iterations that is independent of the problem dimension.

Over the course of the past decade, a variety of randomized algorithms have been proposed for computing approximate least-squares (LS) solutions in large-scale settings. A longstanding practical issue is that, for any given input, the user rarely knows the actual error of an approximate solution (relative to the exact solution). Likewise, it is difficult for the user to know precisely how much computation is needed to achieve the desired error tolerance. Consequently, the user often appeals to worst-case error bounds that tend to offer only qualitative guidance. As a more practical alternative, we propose a bootstrap method to compute a posteriori error estimates for randomized LS algorithms. These estimates permit the user to numerically assess the error of a given solution, and to predict how much work is needed to improve a "preliminary" solution. In addition, we provide theoretical consistency results for the method, which are the first such results in this context (to the best of our knowledge). From a practical standpoint, the method also has considerable flexibility, insofar as it can be applied to several popular sketching algorithms, as well as a variety of error metrics. Moreover, the extra step of error estimation does not add much cost to an underlying sketching algorithm. Finally, we demonstrate the effectiveness of the method with empirical results.

Large-scale distributed optimization is of great importance in various applications. For data-parallel based distributed learning, the inter-node gradient communication often becomes the performance bottleneck. In this paper, we propose the error compensated quantized stochastic gradient descent algorithm to improve the training efficiency. Local gradients are quantized to reduce the communication overhead, and accumulated quantization error is utilized to speed up the convergence. Furthermore, we present theoretical analysis on the convergence behaviour, and demonstrate its advantage over competitors. Extensive experiments indicate that our algorithm can compress gradients by a factor of up to two magnitudes without performance degradation.

The multicategory SVM (MSVM) of Lee et al. (2004) is a natural generalization of the classical, binary support vector machines (SVM). However, its use has been limited by computational difficulties. The simplex-cone SVM (SCSVM) of Mroueh et al. (2012) is a computationally efficient multicategory classifier, but its use has been limited by a seemingly opaque interpretation. We show that MSVM and SCSVM are in fact exactly equivalent, and provide a bijection between their tuning parameters. MSVM may then be entertained as both a natural and computationally efficient multicategory extension of SVM. We further provide a Donsker theorem for finite-dimensional kernel MSVM and partially answer the open question pertaining to the very competitive performance of One-vs-Rest methods against MSVM. Furthermore, we use the derived asymptotic covariance formula to develop an inverse-variance weighted classification rule which improves on the One-vs-Rest approach.

We show that Entropy-SGD (Chaudhari et al., 2017), when viewed as a learning algorithm, optimizes a PAC-Bayes bound on the risk of a Gibbs (posterior) classifier, i.e., a randomized classifier obtained by a risk-sensitive perturbation of the weights of a learned classifier. Entropy-SGD works by optimizing the bound’s prior, violating the hypothesis of the PAC-Bayes theorem that the prior is chosen independently of the data. Indeed, available implementations of Entropy-SGD rapidly obtain zero training error on random labels and the same holds of the Gibbs posterior. In order to obtain a valid generalization bound, we rely on a result showing that data-dependent priors obtained by stochastic gradient Langevin dynamics (SGLD) yield valid PAC-Bayes bounds provided the target distribution of SGLD is eps-differentially private. We observe that test error on MNIST and CIFAR10 falls within the (empirically nonvacuous) risk bounds computed under the assumption that SGLD reaches stationarity. In particular, Entropy-SGLD can be configured to yield relatively tight generalization bounds and still fit real labels, although these same settings do not obtain state-of-the-art performance.

We study active object tracking, where a tracker takes as input the visual observation (i.e. frame sequence) and produces the camera control signal (e.g., move forward, turn left, etc). Conventional methods tackle the tracking and the camera control separately, which is challenging to tune jointly. It also incurs many human efforts for labeling and many expensive trial-and-errors in real-world. To address these issues, we propose, in this paper, an end-to-end solution via deep reinforcement learning, where a ConvNet-LSTM function approximator is adopted for the direct frame-to-action prediction. We further propose an environment augmentation technique and a customized reward function, which are crucial for a successful training. The tracker trained in simulators (ViZDoom, Unreal Engine) shows good generalization in the case of unseen object moving path, unseen object appearance, unseen background, and distracting object. It can restore tracking when occasionally losing the target. With the experiments over the VOT dataset, we also find that the tracking ability, obtained solely from simulators, can potentially transfer to real-world scenarios.

The multivariate probit model (MVP) is a popular classic model for studying binary responses of multiple entities. Nevertheless, the computational challenge of learning the MVP model, given that its likelihood involves integrating over a multidimensional constrained space of latent variables, significantly limits its application in practice. We propose a flexible deep generalization of the classic MVP, the Deep Multivariate Probit Model (DMVP), which is an end-to-end learning scheme that uses an efficient parallel sampling process of the multivariate probit model to exploit GPU-boosted deep neural networks. We present both theoretical and empirical analysis of the convergence behavior of DMVP’s sampling process with respect to the resolution of the correlation structure. We provide convergence guarantees for DMVP and our empirical analysis demonstrates the advantages of DMVP’s sampling compared with standard MCMC-based methods. We also show that when applied to multi-entity modelling problems, which are natural DMVP applications, DMVP trains faster than classical MVP, by at least an order of magnitude, captures rich correlations among entities, and further improves the joint likelihood of entities compared with several competitive models.